Sparse Representation for Information Fusion

Open Access
Author:
Bahrampour, Soheil
Graduate Program:
Electrical Engineering
Degree:
Doctor of Philosophy
Document Type:
Dissertation
Date of Defense:
March 03, 2015
Committee Members:
  • Asok Ray, Dissertation Advisor
  • William Kenneth Jenkins, Committee Chair
  • Shashi Phoha, Committee Member
  • Jeffrey Scott Mayer, Committee Member
  • Joseph Francis Horn, Committee Member
Keywords:
  • Multimodal Dictionary Learning
  • Sparse Representation
  • Information Fusion
Abstract:
Sparse representation methods have recently attracted much attention in the signal processing and machine learning community. The main underlying idea is that most of natural signals are inherently sparse in certain bases or dictionaries. In this dissertation, the problem of information fusion using sparsity models are investigated. For this purpose, different sparsity models, including the joint sparsity and the tree-structured sparsity priors are studied and a flexible framework for extraction of cross-correlated information from different sources is proposed, which allows for fusion of the modalities at multiple granularities. In addition, the proposed method quantifies the quality of the different sources that offers added robustness to several sparsity-based multimodal classification algorithms. While several fixed dictionaries have been suggested for sparse representation in different applications, a better approach would be to train the dictionary for the application in hand, which is referred to as dictionary learning. However, most of the existing dictionary learning algorithms are only applicable with a single source of information. In this dissertation, reconstructive-based and discriminative-based multimodal dictionary learning algorithms are developed, which have several implications. In particular, the algorithms result in more compact dictionaries that are critical for efficient implementation of the sparse models. Moreover, the fusion is performed both at the feature level and the decision level for robust fusion. The efficacy of the proposed algorithms is evaluated on several applications including face recognition, biometrics recognition, and target classification. Kernelized dictionary learning algorithms are also developed for sparse representation in the featured space. In the proposed formulation, the dictionary and classifier are obtained jointly for optimal classification performance which provides the learned features. The proposed algorithms are then used for hyperspectral image classification, where they enhance collaboration among the neighboring pixels and results in the state-of-the-art performance.